GEO Lab #04 · S4-GROK-Q2-R3 · Original response
This page presents one selected response from GEO Lab #04. It adds a smaller reading and citation entry point; it does not replace the full archive or collect a new answer.
The OpenAI arm uses the Responses API with web search and does not represent consumer ChatGPT Search. Grok records identify their actual consumer-interface mode and available model label.
Recorded conditions
Original prompt
What credible experiments have measured whether GEO tactics increase AI brand mentions or recommendations? Include sources when available.
The agreed prompt is reproduced exactly. Any recorded difference in the actual submitted text or search UI is retained in the raw record and full appendix.
Original answer
Credible experiments exist, but the strongest evidence is concentrated on content that is already retrieved or placed in an LLM’s context. Evidence that on-page GEO tactics alone reliably increase organic discoverability (being found and mentioned in the first place) across live engines is thinner and more mixed. A July 2026 critical survey of 45 studies (Martinez, arXiv:2607.14035) concluded that already-retrieved content can causally change how it is cited or used, but no reviewed technique showed a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior. Topical relevance and position in context are among the more reproducible factors. Run-to-run variability in AI answers is high.arXiv Foundational controlled study (visibility once content is in context) Aggarwal et al., “GEO: Generative Engine Optimization” (KDD 2024; arXiv:2311.09735) from researchers affiliated with Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi. This is the peer-reviewed paper that popularized the term. Setup: GEO-bench (10,000 diverse queries + relevant web sources). Nine content-rewrite strategies were applied and evaluated in a generative-engine / RAG-style setup measuring Position-Adjusted Word Count (share of the answer attributed to the source, weighted by position) and Subjective Impression. Results (approximate relative lifts on Position-Adjusted Word Count vs. baseline ~19.3): Quotation Addition ~+41%, Statistics Addition ~+31–33%, Fluency Optimization and Cite Sources ~+28–29%. Keyword stuffing was neutral-to-negative. Gains were largest for lower-ranked sources (e.g., Cite Sources ~+115% for rank-5 sources; top-ranked sources sometimes lost share). Combinations could add modest further gains. Additional check: Methods also improved visibility on Perplexity (via uploaded files) by up to ~37%. Caveat: Primarily tests absorption/citation once sources are already in context, not whether the page gets retrieved from the live web. The widely cited “up to 40%” figure refers to this share-of-answer metric, not clicks, traffic, or organic discovery probability.Tinyurl Live-web / in-the-wild tests of rewriting strategies LIVE-GEO (arXiv:2611.01126, Oct 2026): Published product pages for 80 fictional brands across 18 families on a public site. Each brand had an “Original” product and versions rewritten with five GEO strategies (CORE, C-SEO, E-GEO, AgenticGEO, Statistics Addition). Evaluated on ChatGPT (GPT-6), Gemini, Grok, plus direct-text API evaluators. With frozen ±0.2% prices and product URLs supplied, CORE generally led Top-1 recommendation rates (roughly 66–76% on the three GEs; higher in some equal-price / order conditions). Other strategies varied. GEs agreed on the top product for only ~49–66% of brands. ChatGPT showed order sensitivity. Statistics Addition introduced unsupported numerical claims that a verifier flagged. Takeaway: Certain rewriting strategies can influence ranking/recommendation once pages are accessible and candidates are presented, but results are inconsistent across engines, sensitive to presentation, and do not guarantee identical outcomes on the live web. Fictional brands limit generalizability.arXiv “What Gets Cited: Competitive GEO” (arXiv:2605.25517): Controlled two-document RAG testbed (252,000 trials across six LLMs). Documents differed in exactly one of 18 factors. Topical relevance, list position, explicit price information, and recent timestamps were the strongest drivers of being cited first. Completeness and trust cues helped modestly; pure formatting edits had little impact. Retrieval vs. page features and brand knowledge Chen & Koudas (University of Toronto), “What Drives AI Citations?” (arXiv:2610.15033, Oct 2026): Large observational study (158,099 answers, ~468k citations over 90 days in two markets) plus API replays and controlled tests. Within a website, FAQ blocks, schema.org markup, summary boxes, author boxes, and page speed showed little or no association with citations (they appeared helpful only in across-website comparisons—Simpson’s paradox). Title and paragraph match to the questions people ask did help. Replay through engines’ APIs: Every ChatGPT-cited URL came from its own search. Retrieval explained ~32% of within-website citation variation vs. ~2.8% for page features. Engines rewrite queries; results diverge from Google/Bing rankings. Model priors matter: The model names brands it “knows” even without a supporting cited page (~46% of unbranded mentions in one analysis) and writes known brands into its searches. Controlled tests showed brand names reordered rankings of otherwise identical products. Implication: Being retrieved by the engine’s own search and existing in the model’s knowledge are larger levers than typical on-page checklist items.arXiv Bouet (Roskilde University, 2026): Pre-registered RCT on eight synthetic running-shoe brands (identical sites; four with Schema.org markup, four without). Measured mention rates on Gemini 2.5 Flash with Google search grounding across 28 prompts. Null result (effect near zero; confidence interval ruled out the large gains often claimed by the industry). Low-authority pages were not retrieved highly enough for the markup to reach the model. Practitioner and directional experiments OtterlyAI public experiments (multiple posts, 2026; they track prompts across engines and often use controls or before/after windows): Publication-date update only on 10 articles: Citations rose ~31.5% (4,800 → 6,312) over 30 days while two controls declined ~10.6%. Engine-level results varied (some declined). Award win: Lift concentrated on award-specific and year-tagged prompts (e.g., specified-awards coverage roughly doubled); overall coverage largely flat. Most of the relevant citations traced to the announcement/press sources. New brand from zero: Off-site mentions (directories/authoritative third-party pages) produced measurable AI mentions and share-of-voice within 14 days. Year-in-title test: Apparent lifts were largely explained by one page and a concurrent control rise; authors concluded the effect was smaller than noise. Authors generally describe results as directional and note high natural variability.Otterly Search Engine Land (2026): Two structured tracking experiments on commercial-intent keywords (one multi-month for an existing brand involving PR/guest posts/listicles; one 30-day cold-start for a SaaS agency). Manual checks across multiple platforms. From low/zero baselines, appearances and citations rose (e.g., 298 appearances in 30 days in the cold-start case). Listicles and specific third-party placements were frequent citation sources. Not a fully controlled RCT, but documented with keyword sets and platform breakdowns.Searchengineland Seer Interactive GEO Olympics (2026): Large observational hypothesis testing (~231k LLM responses across 7 platforms over 52 days around the Olympics). Strong “signal architecture” (entity authority such as Wikipedia/official sources + third-party validation + community discussion) was associated with substantially higher mention volume (roughly 7.8× gap between athletes with all three vs. thin stacks). Tests confirmed several hypotheses around narrative persistence, social proof, citation sources by query type, etc. Correlational rather than pure intervention. Bottom line What has been measured positively: Adding quotations, statistics, and credible citations (Princeton) increases a source’s share of an answer when it is already in context. Certain optimized rewrites (LIVE-GEO CORE) can raise Top-1 rates among presented candidates. Query-content alignment and retrieval matter more than many checklist tactics. Off-site mentions, listicles, press, and entity signals are repeatedly linked to mentions in practitioner and observational work. What is weaker or null in better-controlled tests: Schema markup alone on low-authority synthetic brands (Bouet); many FAQ/schema/formatting features once website is held constant (Chen & Koudas); stable cross-platform organic discoverability gains (critical survey). Practical implication: GEO effects are real for how retrieved content is used and for building entity/knowledge presence, but “optimize the page and mentions will rise” is overstated relative to the current evidence. High response variability means experiments need repeated measurements, controls, and clear separation of retrieval vs. citation/absorption stages.
Recorded sources
These are the source roles recorded in the public dataset. A retrieval or source-panel item is not automatically an answer-body citation. A URL supplied in the prompt and repeated by an answer does not prove content access. No destination has been repaired or independently validated here.
Recorded answer-body links (6)
- arXivAnswer-body link/citation · arxiv.org
- GEO: Generative Engine Optimization - Princeton University (Tinyurl)Answer-body link/citation · tinyurl.com
- [2611.01126] Do GEO Gains Survive the Live Web? Evaluating In-the-Wild Visibility Gains of Generative Engine Optimization via LIVE-GEO (arXiv)Answer-body link/citation · arxiv.org
- [2610.15033] What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO) (arXiv)Answer-body link/citation · arxiv.org
- GEO Experiments & AI Search Case Studies | OtterlyAI (Otterly)Answer-body link/citation · otterly.ai
- SearchenginelandAnswer-body link/citation · searchengineland.com
Retrieval-only / source-panel inventory (3)
- searchengineland.com/geo-experiments-challenge-conventional-ai-visibility-advice-488342Retrieval activity, not body citation · searchengineland.com
- arxiv.org/html/2611.01126v1Retrieval activity, not body citation · arxiv.org
- seerinteractive.com/insights/what-ai-thinks-about-your-brand-is-already-writtenRetrieval activity, not body citation · www.seerinteractive.com
Evidence and qualifications
This is the existing redacted public record. It has not been rewritten for this page. Earlier privacy edits and collection qualifications continue to apply. Editorial text is English; a few original captures retain non-English UI badges or source-panel date labels. An answer’s empirical claims are platform output, not verified findings of this archive.
Capture method
Assistant read full native rendered answer; matching prompt checked; DOM body and expanded citation groups captured.
Body review
{
"reviewer": "assistant",
"complete_body_read": true,
"coherent_ending": true,
"inline_url_coverage_checked": true,
"external_claims_verified": false,
"independent_human_double_coding": false
}
Existing public-text privacy policy
{
"body_changed": false,
"policy": "Only actual URL tracking, signed/session URL parameters and necessary identifying information are removed; original wording, including analytics parameter examples outside URLs, remains.",
"annotation_positions": "Original position indexes omitted because URL/privacy edits can change character locations."
}
Prompt SHA-256: 84f18194f3ea8dd08151f46577cb5614af08d775b76bd516de9d149a7fc01118
Answer SHA-256: e501d7dce2de3ca999e3a1b2f3ba7705d12c790681a7e9975a6599f0fb11b286